Paper Review 9: Key-Value Retrieval Networks for Task-Oriented Dialogue

In this post, the paper “Key-Value Retrieval Networks for Task-Oriented Dialogue” is summarized.

Link to paper: http://www.aclweb.org/anthology/W17-5506

Mihail Eric, Lakshmi Krishnan, Francois Charette, Christopher D. Manning, 2017, “Key-Value Retrieval Networks for Task-Oriented Dialogue,” in Proceedings of the SIGDIAL 2017 Conference, pages 37–49

In this paper, researchers from Stanford NLP Group and Ford Research and Innovation Center seek to address this problem by proposing a new neural dialogue agent that is able to effectively sustain grounded, multi-domain discourse through a novel key-value retrieval mechanism. Moreover, they release a new dataset of 3,031 dialogues covering three distinct tasks in the in-car personal assistant space: calendar scheduling, weather information retrieval, and point-of-interest navigation.

Their model starts with an encoder-decoder sequence architecture and is further augmented with an attention-based retrieval mechanism that effectively reasons over a key-value representation of the underlying knowledge base. The data for the multi-turn dialogues was collected using a Wizard-of-Oz scheme. 241 unique workers from Amazon Mechanical Turk were anonymously recruited to use the interface based on this scheme they built over a period of about six days.

To compare their methods, they employ a rule-based model that with modular dialogue state trackers, KB query, and natural language generation components, and a copy-augmented sequence-to-sequence network and its variant with encoder attention relying solely on dialogue history for system response generation. Experiments show that their model outperforms competitive heuristic and neural baselines on both automatic and human evaluation metrics.